Forecasting Electricity Prices: Autoregressive Hybrid Nearest Neighbors (ARHNN) Method

Abstract

The ongoing reshape of electricity markets has significantly stimulated electricity trading. Limitations in storing electricity as well as on-the-fly changes in demand and supply dynamics, have led price forecasts to be a fundamental aspect of traders' economic stability and growth. In this perspective, there is a broad literature that focuses on developing methods and techniques to forecast electricity prices. In this paper, we develop a new hybrid method, called ARHNN, for electricity price forecasting (EPF) in day-ahead markets. A well performing autoregressive model, with exogenous variables, is the main forecasting instrument in our method. Contrarily to the traditional statistical approaches, in which the calibration sample consists of the most recent and successive observations, we employ the k-nearest neighbors (k-NN) instance-based learning algorithm and we select the calibration sample based on a similarity (distance) measure over a subset of the autoregressive model's variables. The optimal levels of the k-NN parameter are identified during the validation period in a way that the forecasting error is minimized. We apply our method in the EPEX SPOT market in Germany. The results reveal a significant improvement in accuracy compared to commonly used approaches.

Calibration window selection based on change-point detection for forecasting electricity prices

“…For comparison purposes, we use the same dataset as in [7]. It spans six years (2015-2020) at hourly resolution and includes four series from the German EPEX SPOT market: electricity spot prices P d,h (more precisely: prices set in the dayahead auction on day d − 1 for the 24h of day d) and day-ahead load Ld,h , wind Ŵd,h and solar power generation Ŝd,h forecasts, see Fig.…”

“…1. The first two years are exclusively used for estimating the Autoregressive Hybrid Nearest Neighbors (ARHNN) method [7]; the remaining methods require less data for calibration. The last 736 days constitute the out-of-sample test period.…”

“…Note that in [7] a different transformation was used. All series, not just P d,h , were normalized by subtracting the mean and dividing by the standard deviation in each calibration window.…”

“…All series, not just P d,h , were normalized by subtracting the mean and dividing by the standard deviation in each calibration window. We denote models utilizing asinh-transformed data with subscript H; the remaining ones use the standard normalization, as in [7].…”

“…This work is inspired by a recent article [2], which utilized a relatively simple change-point detection method [9] to split the time series into segments with the 'same' price level, and an ICCS 2021 paper [7], which employed the k-nearest neighbors (k-NN) algorithm to select the calibration sample based on similarity over a subset of explanatory variables. Here, we utilize a recently proposed change-point detection algorithm -the Narrowest-Over-Threshold (NOT) Fig.…”